Add __assume builtin function

README.md

@@ -23,7 +23,7 @@ The goal is to implement the actual C standard and not some subset for performan
 
 ## Limits and Errors
 
-There are still several open areas, but most targets have been reached.  The current Dhrystone performance is 61 iterations per second with byte code (11434) and 270 iterations with native code (12145 Bytes).
+There are still several open areas, but most targets have been reached.  The current Dhrystone performance is 61 iterations per second with byte code (11504) and 278 iterations with native code (12184 Bytes).
 
 ### Language
 
@@ -34,7 +34,7 @@ There are still several open areas, but most targets have been reached.  The cur
 
 ### Standard Libraries
 
-* No file functions
+* No standard file functions, but CBM based file ops
 
 ### Runtime
 
@@ -44,7 +44,6 @@ There are still several open areas, but most targets have been reached.  The cur
 ### Optimizing
 
 * Complex loop optimization
-* Partial block domination analysis
 * Auto variables placed on fixed stack for known call sequence
 
 ### Intermediate code generation
@@ -54,15 +53,13 @@ There are still several open areas, but most targets have been reached.  The cur
 
 ### Native code generation
 
-* More byte operation optimisation required
-* Simple loop detection and optimisation not complete
-
 ## Compiler arguments
 
 The compiler is command line driven, and creates an executable .prg file.
 
     oscar64 {-i=includePath} [-o=output.prg] [-rt=runtime.c] [-e] [-n] [-dSYMBOL[=value]] {source.c}
     
+* -v : verbose output for diagnostics
 * -i : additional include paths
 * -o : optional output file name
 * -rt : alternative runtime library, replaces the crt.c
@@ -91,6 +88,8 @@ Will switch to the lowercase PETSCII font and translate the strings while printi
 PETSCII string literals can also be generated using a "p" or "P" prefix such as:
 
 	printf(p"Hello World\n");
+	
+Screen codes can be generated similar using "s" or "S" prefix.
 
 Input from the console will also be translated accordingly.
 
@@ -113,6 +112,13 @@ A section of the file can be selected by providing a limit and or an offset into
 	};
 
 
+## Language extensions for optimization
+
+### Additional Optimizer information using __assume()
+
+The compiler can be provided with additional information using the built in function __assume(cond).  This can be useful to mark unreachable code using __assume(false) for e.g. the default of a switch statement.  Another good option is to limit the value range of arguments to allow the compiler using byte operations without the need for integer promotion.
+
+
 
 ## Inline Assembler
 

include/c64/sprites.c

@@ -12,6 +12,8 @@ void spr_init(char * screen)
 
 void spr_set(char sp, bool show, int xpos, int ypos, char image, char color, bool multi, bool xexpand, bool yexpand)
 {
+	__assume (sp < 8);
+
 	char	m = 1 << sp;
 
 	if (show)
@@ -47,6 +49,8 @@ void spr_set(char sp, bool show, int xpos, int ypos, char image, char color, boo
 
 void spr_show(char sp, bool show)
 {
+	__assume (sp < 8);
+
 	if (show)
 		vic.spr_enable |= 1 << sp;
 	else
@@ -55,6 +59,8 @@ void spr_show(char sp, bool show)
 
 void spr_move(char sp, int xpos, int ypos)
 {
+	__assume (sp < 8);
+
 	vic.spr_pos[sp].y = ypos;
 	vic.spr_pos[sp].x = xpos & 0xff;
 	if (xpos & 0x100)
@@ -65,11 +71,15 @@ void spr_move(char sp, int xpos, int ypos)
 
 void spr_image(char sp, char image)
 {
+	__assume (sp < 8);
+
 	vspriteScreen[sp] = image;
 }
 
 void spr_color(char sp, char color)
 {
+	__assume (sp < 8);
+
 	vic.spr_color[sp] = color;
 }
 

oscar64/ByteCodeGenerator.cpp

@@ -4250,6 +4250,9 @@ void ByteCodeBasicBlock::Compile(InterCodeProcedure* iproc, ByteCodeProcedure* p
 			}
 			return;
 
+		case IC_UNREACHABLE:
+			this->Close(proc->exitBlock, nullptr, BC_JUMPS);
+			return;
 		}
 
 		i++;

oscar64/Declaration.h

@@ -133,7 +133,8 @@ enum ExpressionType
 	EX_SWITCH,
 	EX_CASE,
 	EX_DEFAULT,
-	EX_CONDITIONAL
+	EX_CONDITIONAL,
+	EX_ASSUME
 };
 
 class Expression

oscar64/GlobalAnalyzer.cpp

@@ -352,6 +352,7 @@ Declaration * GlobalAnalyzer::Analyze(Expression* exp, Declaration* procDec)
 		break;
 	case EX_BREAK:
 	case EX_CONTINUE:
+	case EX_ASSUME:
 		break;
 	case EX_TYPE:
 		break;

oscar64/InterCode.cpp

@@ -2501,6 +2501,9 @@ void InterInstruction::Disassemble(FILE* file)
 		case IC_RETURN:
 			fprintf(file, "RET");
 			break;
+		case IC_UNREACHABLE:
+			fprintf(file, "UNREACHABLE");
+			break;
 		}
 		static char typechars[] = "NBCILFP";
 
@@ -2562,6 +2565,7 @@ InterCodeBasicBlock::InterCodeBasicBlock(void)
 	mLoopHead = false;
 	mChecked = false;
 	mTraceIndex = -1;
+	mUnreachable = false;
 }
 
 InterCodeBasicBlock::~InterCodeBasicBlock(void)
@@ -5371,7 +5375,7 @@ bool InterCodeBasicBlock::PushSinglePathResultInstructions(void)
 				if (ins->mDst.mTemp >= 0 && !providedTemps[ins->mDst.mTemp] && !requiredTemps[ins->mDst.mTemp])
 				{
 					int j = 0;
-					while (j < ins->mNumOperands && !(ins->mSrc[j].mTemp >= 0 && providedTemps[ins->mSrc[j].mTemp]) && !IsTempModifiedOnPath(ins->mSrc[j].mTemp, i + 1))
+					while (j < ins->mNumOperands && (ins->mSrc[j].mTemp < 0 || !(providedTemps[ins->mSrc[j].mTemp] || IsTempModifiedOnPath(ins->mSrc[j].mTemp, i + 1))))
 						j++;
 
 					if (j == ins->mNumOperands && IsMoveable(ins->mCode) && (ins->mCode != IC_LOAD || !hadStore))
@@ -5741,6 +5745,66 @@ bool InterCodeBasicBlock::IsEqual(const InterCodeBasicBlock* block) const
 	return false;
 }
 
+bool InterCodeBasicBlock::DropUnreachable(void)
+{
+	if (!mVisited)
+	{
+		mVisited = true;
+
+		int i = 0;
+		while (i < mInstructions.Size() && mInstructions[i]->mCode != IC_UNREACHABLE)
+			i++;
+
+		if (i < mInstructions.Size())
+		{
+			// kill all instructions after this
+			mInstructions.SetSize(i + 1);
+			mFalseJump = nullptr;
+			mTrueJump = nullptr;
+
+			if (mInstructions.Size() == 1)
+				mUnreachable = true;
+		}
+		else
+		{
+			if (mFalseJump)
+			{
+				if (mFalseJump->DropUnreachable())
+				{
+					mInstructions.Last()->mCode = IC_JUMP;
+					mInstructions.Last()->mNumOperands = 0;
+					mFalseJump = nullptr;
+
+					if (mTrueJump->DropUnreachable())
+					{
+						mTrueJump = nullptr;
+						mInstructions.SetSize(mInstructions.Size() - 1);
+						if (mInstructions.Size() == 0)
+							mUnreachable = true;
+					}
+				}
+				else if (mTrueJump->DropUnreachable())
+				{
+					mInstructions.Last()->mCode = IC_JUMP;
+					mInstructions.Last()->mNumOperands = 0;
+					mTrueJump = mFalseJump;
+					mFalseJump = nullptr;
+				}
+			}
+			else if (mTrueJump && mTrueJump->DropUnreachable())
+			{
+				mTrueJump = nullptr;
+				mInstructions.SetSize(mInstructions.Size() - 1);
+				if (mInstructions.Size() == 0)
+					mUnreachable = true;
+			}
+		}
+	}
+
+	return mUnreachable;
+}
+
+
 bool InterCodeBasicBlock::OptimizeIntervalCompare(void)
 {
 	bool	changed = false;
@@ -7480,6 +7544,21 @@ void InterCodeProcedure::Close(void)
 
 	DisassembleDebug("Simplified range limited relational ops");
 #endif
+#if 1
+	ResetVisited();
+	mEntryBlock->DropUnreachable();
+
+	ResetEntryBlocks();
+	ResetVisited();
+	mEntryBlock->CollectEntryBlocks(nullptr);
+
+	BuildDataFlowSets();
+	TempForwarding();
+	RemoveUnusedInstructions();
+	
+	DisassembleDebug("Removed unreachable branches");
+#endif
+
 	MapVariables();
 
 	DisassembleDebug("mapped variabled");

oscar64/InterCode.h

@@ -35,7 +35,8 @@ enum InterCode
 	IC_RETURN_STRUCT,
 	IC_RETURN,
 	IC_ASSEMBLER,
-	IC_JUMPF
+	IC_JUMPF,
+	IC_UNREACHABLE
 };
 
 enum InterType
@@ -448,7 +449,7 @@ public:
 	InterCodeBasicBlock			*	mTrueJump, * mFalseJump, * mDominator;
 	GrowingInstructionArray			mInstructions;
 
-	bool							mVisited, mInPath, mLoopHead, mChecked, mConditionBlockTrue;
+	bool							mVisited, mInPath, mLoopHead, mChecked, mConditionBlockTrue, mUnreachable;
 
 	NumberSet						mLocalRequiredTemps, mLocalProvidedTemps;
 	NumberSet						mEntryRequiredTemps, mEntryProvidedTemps;
@@ -572,6 +573,8 @@ public:
 
 	void CompactInstructions(void);
 	bool OptimizeIntervalCompare(void);
+
+	bool DropUnreachable(void);
 };
 
 class InterCodeModule;

oscar64/InterCodeGenerator.cpp

@@ -2605,6 +2605,25 @@ InterCodeGenerator::ExValue InterCodeGenerator::TranslateExpression(Declaration*
 			return ExValue(TheVoidTypeDeclaration);
 		}
 
+		case EX_ASSUME:
+		{
+#if 1
+			InterCodeBasicBlock* tblock = new InterCodeBasicBlock();
+			proc->Append(tblock);
+			InterCodeBasicBlock* fblock = new InterCodeBasicBlock();
+			proc->Append(fblock);
+
+			TranslateLogic(procType, proc, block, tblock, fblock, exp->mLeft, inlineMapper);
+
+			InterInstruction* ins = new InterInstruction();
+			ins->mCode = IC_UNREACHABLE;
+			fblock->Append(ins);
+			fblock->Close(nullptr, nullptr);
+
+			block = tblock;
+#endif
+			return ExValue(TheVoidTypeDeclaration);
+		}
 		case EX_LOGICAL_NOT:
 		{
 			vl = TranslateExpression(procType, proc, block, exp->mLeft, breakBlock, continueBlock, inlineMapper);

oscar64/NativeCodeGenerator.cpp

@@ -6447,7 +6447,8 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 				int	l = Binlog(ins->mSrc[1].mIntConst & 0xffff);
 
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
-				mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
+				if (!ins->mSrc[0].IsUByte() || ins->mSrc[0].mRange.mMaxValue > 15)
+					mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
 				mIns.Push(NativeCodeInstruction(ASMIT_TAX, ASMIM_IMPLIED));
 
 				NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("bitshift")));
@@ -6466,7 +6467,8 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 				NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
 
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
-				mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
+				if (!ins->mSrc[0].IsUByte() || ins->mSrc[0].mRange.mMaxValue > 15)
+					mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
 				mIns.Push(NativeCodeInstruction(ASMIT_TAX, ASMIM_IMPLIED));
 
 				if (ins->mSrc[1].mTemp < 0)
@@ -6570,7 +6572,8 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 				int	l = Binlog(ins->mSrc[1].mIntConst & 0xffff);
 
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
-				mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
+				if (!ins->mSrc[0].IsUByte() || ins->mSrc[0].mRange.mMaxValue > 15)
+					mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
 				mIns.Push(NativeCodeInstruction(ASMIT_TAX, ASMIM_IMPLIED));
 
 				NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("bitshift")));
@@ -6586,7 +6589,8 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 				NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
 
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
-				mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
+				if (!ins->mSrc[0].IsUByte() || ins->mSrc[0].mRange.mMaxValue > 15)
+					mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
 				mIns.Push(NativeCodeInstruction(ASMIT_TAX, ASMIM_IMPLIED));
 
 				if (ins->mSrc[1].mTemp < 0)
@@ -6730,7 +6734,8 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 				int	l = Binlog(ins->mSrc[1].mIntConst & 0xffff);
 
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
-				mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
+				if (!ins->mSrc[0].IsUByte() || ins->mSrc[0].mRange.mMaxValue > 15)
+					mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
 				mIns.Push(NativeCodeInstruction(ASMIT_TAX, ASMIM_IMPLIED));
 
 				NativeCodeGenerator::Runtime& frt(nproc->mGenerator->ResolveRuntime(Ident::Unique("bitshift")));
@@ -6759,7 +6764,8 @@ NativeCodeBasicBlock* NativeCodeBasicBlock::BinaryOperator(InterCodeProcedure* p
 				NativeCodeBasicBlock* eblock = nproc->AllocateBlock();
 
 				mIns.Push(NativeCodeInstruction(ASMIT_LDA, ASMIM_ZERO_PAGE, BC_REG_TMP + proc->mTempOffset[ins->mSrc[0].mTemp]));
-				mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
+				if (!ins->mSrc[0].IsUByte() || ins->mSrc[0].mRange.mMaxValue > 15)
+					mIns.Push(NativeCodeInstruction(ASMIT_AND, ASMIM_IMMEDIATE, 0x0f));
 				mIns.Push(NativeCodeInstruction(ASMIT_TAX, ASMIM_IMPLIED));
 
 				if (ins->mSrc[1].mTemp < 0)
@@ -9671,6 +9677,39 @@ bool NativeCodeBasicBlock::MoveIndirectLoadStoreUp(int at)
 	return false;
 }
 
+bool NativeCodeBasicBlock::MoveLoadStoreOutOfXYRangeUp(int at)
+{
+	int	j = at - 1;
+	while (j >= 0)
+	{
+		if (mIns[j].MayBeChangedOnAddress(mIns[at + 2]))
+			return false;
+		if (mIns[j].ChangesAddress() && mIns[j].SameEffectiveAddress(mIns[at + 1]))
+			return false;
+
+		if (mIns[j].mType == ASMIT_LDA)
+		{
+			if (j > 0 && (mIns[j - 1].mType == ASMIT_CLC || mIns[j - 1].mType == ASMIT_SEC))
+				j--;
+			mIns.Insert(j, mIns[at + 2]);
+			mIns.Insert(j, mIns[at + 2]);
+			if (j > 0)
+			{
+				mIns[j].mLive |= mIns[j - 1].mLive;
+				mIns[j + 1].mLive |= mIns[j - 1].mLive;
+			}
+			mIns.Remove(at + 3);
+			mIns.Remove(at + 3);
+
+			return true;
+		}
+
+		j--;
+	}
+
+	return false;
+}
+
 bool NativeCodeBasicBlock::MoveAbsoluteLoadStoreUp(int at)
 {
 	int	j = at - 1;
@@ -11702,6 +11741,20 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 
 #endif
 
+#if 1
+		for (int i = 0; i + 2 < mIns.Size(); i++)
+		{
+			if ((mIns[i + 0].mType == ASMIT_TAX || mIns[i + 0].mType == ASMIT_TAY) &&
+				mIns[i + 1].mType == ASMIT_LDA && (mIns[i + 1].mMode == ASMIM_ABSOLUTE || mIns[i + 1].mMode == ASMIM_ZERO_PAGE) &&
+				mIns[i + 2].mType == ASMIT_STA && mIns[i + 2].mMode == ASMIM_ZERO_PAGE && !(mIns[i + 2].mLive & (LIVE_CPU_REG_Z | LIVE_CPU_REG_A)))
+			{
+				if (MoveLoadStoreOutOfXYRangeUp(i))
+					changed = true;
+			}
+		}
+
+#endif
+
 #if 1
 		if (pass > 1)
 		{
@@ -13056,6 +13109,22 @@ bool NativeCodeBasicBlock::PeepHoleOptimizer(int pass)
 						progress = true;
 					}
 #endif
+#if 1
+					else if (
+						mIns[i + 0].mType == ASMIT_STA && mIns[i + 0].mMode == ASMIM_ZERO_PAGE &&
+						mIns[i + 1].mType == ASMIT_LDA && (mIns[i + 1].mMode == ASMIM_ZERO_PAGE && mIns[i + 0].mAddress != mIns[i + 2].mAddress || mIns[i + 1].mMode == ASMIM_ABSOLUTE) &&
+						mIns[i + 2].mType == ASMIT_CLC &&
+						mIns[i + 3].mType == ASMIT_ADC && mIns[i + 3].mMode == ASMIM_ZERO_PAGE && mIns[i + 0].mAddress == mIns[i + 3].mAddress && !(mIns[i + 3].mLive & LIVE_MEM))
+					{
+						mIns[i + 3] = mIns[i + 1];
+						mIns[i + 3].mType = ASMIT_ADC;
+
+						mIns[i + 0].mType = ASMIT_NOP; mIns[i + 0].mMode = ASMIM_IMPLIED;
+						mIns[i + 1].mType = ASMIT_NOP; mIns[i + 1].mMode = ASMIM_IMPLIED;
+
+						progress = true;
+					}
+#endif
 #if 1
 					else if (
 						mIns[i + 0].mType == ASMIT_LDA && mIns[i + 0].mMode == ASMIM_IMMEDIATE &&
@@ -14942,6 +15011,9 @@ void NativeCodeProcedure::CompileInterBlock(InterCodeProcedure* iproc, InterCode
 			}
 			return;
 
+		case IC_UNREACHABLE:
+			block->Close(mExitBlock, nullptr, ASMIT_JMP);
+			return;
 		}
 
 		i++;

oscar64/NativeCodeGenerator.h

@@ -203,6 +203,7 @@ public:
 
 	bool MoveIndirectLoadStoreUp(int at);
 	bool MoveAbsoluteLoadStoreUp(int at);
+	bool MoveLoadStoreOutOfXYRangeUp(int at);
 	bool MoveLoadAddImmStoreUp(int at);
 	bool MoveCLCLoadAddZPStoreUp(int at);
 	bool MoveLoadAddZPStoreUp(int at);

oscar64/Parser.cpp

@@ -1920,6 +1920,13 @@ Expression* Parser::ParseStatement(void)
 				exp = new Expression(mScanner->mLocation, EX_VOID);
 			}
 			break;
+		case TK_ASSUME:
+			mScanner->NextToken();
+			exp = new Expression(mScanner->mLocation, EX_ASSUME);
+			exp->mLeft = ParseParenthesisExpression();
+			break;
+
+
 		default:
 			exp = ParseExpression();
 		}

oscar64/Scanner.cpp

@@ -47,6 +47,7 @@ const char* TokenNames[] =
 	"'static'",
 	"'extern'",
 	"'inline'",
+	"'__assume'",
 
 	"__asm",
 
@@ -1222,6 +1223,8 @@ void Scanner::NextRawToken(void)
 					mToken = TK_INLINE;
 				else if (!strcmp(tkident, "__asm"))
 					mToken = TK_ASM;
+				else if (!strcmp(tkident, "__assume"))
+					mToken = TK_ASSUME;
 				else
 				{
 					mToken = TK_IDENT;

oscar64/Scanner.h

@@ -45,6 +45,7 @@ enum Token
 	TK_STATIC,
 	TK_EXTERN,
 	TK_INLINE,
+	TK_ASSUME,
 
 	TK_ASM,